Best possible solution:

Land this patch or an equivalent narrow fix after exact-head validation passes, keeping the legacy symlink exception scoped to OpenClaw-owned runtime-deps migration and leaving arbitrary symlink replacement blocked.

Do we have a high-confidence way to reproduce the issue?

Yes. A focused reproduction is to make dist/extensions/<plugin>/node_modules a symlink to .local/bundled-plugin-runtime-deps/<id>/node_modules and run bundled runtime-deps staging; current main reaches the generic symlink refusal before materializing replacement deps.

Is this the best way to solve the issue?

Yes. The PR is the narrow maintainable approach because it unlinks only legacy runtime-deps symlinks resolving under the OpenClaw-owned root, then lets the existing replacement/removal and arbitrary-symlink guards continue to enforce the unsafe cases.

Acceptance criteria:

• pnpm test test/scripts/stage-bundled-plugin-runtime-deps.test.ts
• pnpm build
• pnpm check:changed

What I checked:

• Current main symlink guard: assertPathIsNotSymlink throws for any symlinked target, and replaceDirAtomically calls it before replacing runtime deps. (scripts/lib/bundled-runtime-deps-stage-state.mjs:85, e8f9c3e6dedc)
• Installed-root staging hits the guard: stageInstalledRootRuntimeDeps calls the symlink guard before removing an empty runtime-deps tree and reaches replaceDirAtomically before replacing one, so current main blocks legacy symlinked node_modules. (scripts/lib/bundled-runtime-deps-materialize.mjs:158, e8f9c3e6dedc)
• Fallback install staging also hits the guard: The fallback installer replaces staged node_modules through replaceDirAtomically and refuses symlinked removal paths, with no legacy runtime-deps exception on current main. (scripts/stage-bundled-plugin-runtime-deps.mjs:324, e8f9c3e6dedc)
• Existing unsafe-symlink coverage: Current tests already assert that an arbitrary symlinked plugin node_modules directory is refused, which matches the PR's stated safety boundary. (test/scripts/stage-bundled-plugin-runtime-deps.test.ts:684, e8f9c3e6dedc)
• PR diff scope: The supplied PR diff adds removeLegacyBundledRuntimeDepsSymlink, calls it immediately before the existing replacement/removal paths, adds a migration test for OpenClaw-owned legacy symlinks, and adds a single-line changelog fix entry. (scripts/lib/bundled-runtime-deps-stage-state.mjs:95, df8b0436669d)
• Related runtime-deps PR is adjacent, not superseding: The related open fix: simplify bundled runtime dependency repair #75183 body explicitly lists this PR as a nearby runtime-deps lifecycle issue with a different root cause, so it is not a duplicate close reason. (98178132fc5c)

Likely related people:

• steipete: Current-main blame for assertPathIsNotSymlink, replaceDirAtomically, and both script staging call sites points to Peter Steinberger; he also owns the adjacent open runtime-deps repair PR fix: simplify bundled runtime dependency repair #75183 in the supplied context. (role: current staging helper owner and adjacent runtime-deps maintainer; confidence: high; commits: 48794b9f88c3, 3c4851037b6c; files: scripts/lib/bundled-runtime-deps-stage-state.mjs, scripts/lib/bundled-runtime-deps-materialize.mjs, scripts/stage-bundled-plugin-runtime-deps.mjs)
• vincentkoc: Current changelog credits @vincentkoc for recent runtime-deps stage-key work in fix(plugins): canonicalize packageRoot before hashing runtime-deps stage key #75048, and local history shows the associated commit touching bundled runtime-deps behavior adjacent to this migration path. (role: recent adjacent runtime-deps maintainer; confidence: medium; commits: 4b98f0952934; files: src/plugins/bundled-runtime-deps.test.ts, src/plugins/bundled-runtime-deps.ts, CHANGELOG.md)

Remaining risk / open question:

• Exact-head CI and pnpm check:changed results were not available from the provided context, and this read-only review did not run tests or builds.
• The PR body reports a raw Vitest invocation; before merge, validation should use the repo wrapper command for the focused test plus the normal changed gate.

Codex review notes: model gpt-5.5, reasoning high; reviewed against e8f9c3e6dedc.